Design principles
Twelve principles behind the patterns, each with a warning sign: what it looks like when a design breaks it.
1Visibility of system status6 patterns
Every AI result shows its current status and which mode produced it, at a glance, not on click-through. A reviewer who can't see the system's state can't oversee it.
Warning sign Status is only visible inside an item, not on the list or dashboard; the mode isn't labelled.
Patterns: AI Disclosure · Provenance Disclosure · Data Freshness Signal · Fallback Mechanism · Audit Trail · Monitoring Dashboard
2Recognition over recall3 patterns
Evidence sits next to the claim it supports. If people have to navigate away to verify, they won't, and review becomes a rubber stamp.
Warning sign Evidence exists but isn't shown inline; the reviewer has to open a separate view to check it.
Patterns: Inline Evidence · Data Freshness Signal · Hard Gate
3Forcing functions, not warnings2 patterns
The AI never assigns, sends or finalizes a result without explicit human confirmation. A dismissible warning is not a forcing function. This is a system constraint, not a preference.
Warning sign Confirmation is optional, dismissible, or can be bypassed through configuration or the API.
Patterns: Hard Gate · Decouple Review Actions
4Plain language over raw scores2 patterns
Confidence is expressed the way the reviewer would say it in conversation, not as a raw probability. This applies to internal review tools and to anything end users see.
Warning sign Raw scores appear without plain-language framing or context.
Patterns: Confidence Translation · Accessible Communication
5Separate indicators, never blended1 pattern
When the system assesses several dimensions, each gets its own indicator. A single blended score encourages automation bias and prevents real review of each dimension.
Warning sign One combined score; dimensions are aggregated before the reviewer sees them.
Patterns: Separated Confidence Scores
6Progressive disclosure5 patterns
The main view stays scannable. Detail and evidence are available on expand or hover, not all shown by default. Governance information that's available but not forced onto every pixel gets checked more often.
Warning sign Full detail is always visible; people have to scroll past irrelevant information to reach the decision.
Patterns: Provenance Disclosure · Confidence Translation · Inline Evidence · Accessible Communication · Monitoring Dashboard
7Consistent flagged state across surfaces1 pattern
An unresolved issue looks the same in every interface that shows it. Inconsistent treatment across surfaces lets issues slip through the gaps.
Warning sign The same flag uses different labels, visuals or severity levels in different interfaces.
Patterns: Monitoring Dashboard
8Reversibility as a first-class state2 patterns
The system stays overridable after an action, not just preventable before it. Correction and re-approval are requirements, not edge cases. Designing only for prevention assumes the gate never fails.
Warning sign No correction or re-approval path after the first gate; overriding needs a workaround instead of a designed flow.
Patterns: User Action Recovery · Fallback Mechanism
9Capture the reason for overrides3 patterns
When a reviewer overrides or corrects an AI result, the system captures the reason with a lightweight input. That signal improves the system; without it, corrections happen, but nothing changes.
Warning sign Overrides are possible, but there's no reason field; corrections aren't logged or routed back to calibration.
Patterns: User Action Recovery · Audit Trail · Decouple Review Actions
10Two audiences, two oversight surfaces2 patterns
The same flagged issue needs different framing for each audience. Reviewers need tools to resolve it. The affected person needs a clear, appropriately scoped explanation, not the mechanics.
Warning sign The reviewer's view and the end user's view share the same component or copy.
Patterns: Multi-Audience Split · Multi-Audience Explanation
11Earn autonomy incrementally0 patterns
Turning on or raising autonomy carries more friction than lower-risk settings. Reducing human involvement is the highest-leverage governance decision in the system, and the interface should reflect that weight.
Warning sign The autonomy threshold is a one-click toggle with no approval, no evidence and no more friction than a minor setting.
Patterns: Patterns coming soon
12Explain to fit the decision5 patterns
How much to explain depends on the decision at hand: the question is when, to whom and in how much detail. Over-explaining causes its own fatigue and teaches people to ignore governance information.
Warning sign The same level of detail appears regardless of role, stakes or stage in the flow.
Patterns: Confidence Translation · Multi-Audience Split · Accessible Communication · Multi-Audience Explanation · Hard Gate